Recent years, Generative Adversarial Networks (GANs) have achieved tremendous success in image synthesis, which usually employ the convolutional operation to extract image features. However, most existing convolutional GANs only extract features in a local neighborhood at a time, which may often cause a lack of non-local information resulting in generating the wrong semantic object in the wrong position. In this paper, we propose a Graph Convolutional Architecture (GCA) for GANs to tackle this problem. GCA constructs a pixel-level graph structure between image regions through an attention mechanism and leverages Graph Convolutional Networks (GCNs) to extract non-local features. GCA extracts the connections between different regions of the image through GCNs, which is a more effective method of using relationship information than directly adding long-range dependencies to the model. We implement the GCA into Deep Convolutional Generative Adversarial Networks (DCGAN), Self-Attention Generative Adversarial Networks (SAGAN), and Concurrent-Single-Image-GAN (ConSinGAN). Extensive experiments are conducted to verify the performance of GCA. The results demonstrate that the GCA can significantly boost the quality of the generated image with more non-local features.